In practice : if I understand well, your lib would first need to download 5 Mb, once and for all, and would then be used as we use Fuse.js right now ?

How well does it handle languages other than English ?

Could it be "trained" on the OpenStreetMap tag wiki ?

Thanks a lot for your work.

See the `embeddings` branch where I tried ternlight. Not convinced, but I'm maybe doing something wrong. Also see if responses to https://news.ycombinator.com/item?id=48814963 > Awesome. We have a dictionary of words to OpenStreetMap tags here : https://codeberg.org/cartes/web/src/branch/master/components... Do you think your work could help us let users type "pancake" and get "crêpe" without writing an explicit "pancake = crêpe" dictionary entry ? In practice : if I understand well, your lib would first need to download 5 Mb, once and for all, and would then be used as we use Fuse.js right now ? How well does it handle languages other than English ? Could it be "trained" on the OpenStreetMap tag wiki ? Thanks a lot for your work.

laem commented

2026年07月07日 13:48:26 +02:00

Author

Owner

Copy link

Commentaire intéressant.

The OSM tag wiki would probably not make for good training data because it only has a single short description for each tag (and even on the French wiki many descriptions seem to be in English?) whereas you want to map multiple descriptions to the same tag.
Ideally you would have real query data (e.g. from cartes.app telemetry), then you could get a LLM to write a bespoke Overpass query for each one and use that as the ground truth. Alternatively, start from the list of OSM tag values used in the wild and ask an LLM to list possible reasons to visit that POI.
You could then use that data to finetune an embedding model for your use case. But, you know, somewhere in that model there's going to be a token vocabulary that the model knows about and at the other end you get a similarity score for each tag value. If you don't need to support complex queries where interactions between words matter ("any restaurant that is NOT Korean"), you could get away with a simple list of words and tags that they match to. Which is right where you started, except it could be more exhaustive. Why limit yourself to two Korean dishes when you can have a LLM list many more for you?

Commentaire intéressant. ``` The OSM tag wiki would probably not make for good training data because it only has a single short description for each tag (and even on the French wiki many descriptions seem to be in English?) whereas you want to map multiple descriptions to the same tag. Ideally you would have real query data (e.g. from cartes.app telemetry), then you could get a LLM to write a bespoke Overpass query for each one and use that as the ground truth. Alternatively, start from the list of OSM tag values used in the wild and ask an LLM to list possible reasons to visit that POI. You could then use that data to finetune an embedding model for your use case. But, you know, somewhere in that model there's going to be a token vocabulary that the model knows about and at the other end you get a similarity score for each tag value. If you don't need to support complex queries where interactions between words matter ("any restaurant that is NOT Korean"), you could get away with a simple list of words and tags that they match to. Which is right where you started, except it could be more exhaustive. Why limit yourself to two Korean dishes when you can have a LLM list many more for you? ```

No Branch/Tag specified

Branches Tags

master

chore/i18n-missing-strings-step-3

abostats

SynchroM-master

gps

embeddings

lysios-patch-5

lysios-patch-4

lysios-patch-3

lysios-patch-2

lysios-patch-1

abo-serveur2

clim

zone-en-favori

categories_toujours_visibles

rechercher_ici_avant_resultats

test-2365

panoramax-contribution

add-categories-validations

gel-bug-prod

osm-note-2

broken

wikidata_booking

fuse7.3

lysios-patrimoine-3

gps2

atproto-fix-2

styles-secondaires-mais-maison

atproto-scopes

itineraire-printemps-2026

modal-sheet-5.6

nlnet-cost-climate-perf

pb-icones-religieux

couvoiture

etiennejr-patch-2

meteo-fainéante

logo-static-url

poc-page-vitrine

remove-compass

test-reproduce-1795

display-cycleway-lanes-2

poc-image

add-media-condition

mapterhorn

test-indoor-equal

outdoors-VS-indoorEqual

hiking-colors

fix-blog-typing

poc-next-international

fix-map-style-modal-not-closing

i18n

hide-odd-zoom-levels

filter_categories

resume_wikipedia

deplacement_bouttons

fix-1095

mobile_permettre_selection

position-carte-initiale

dégradé-icônes

better_wikipedia

affichage_donnees_wikidata

mobile-permettre-selection

avis-details

perf-seo

modal-sheet-5

plans-novembre

display-cycleway-lanes

fix-1227

fix-1234

etiennejr-patch-1

update-1240

fix-1236

fix-1131

photon-search-street

fix-1221

article-avis

net-haut-zoom

baies-moins-visibles

belles-fiches-lieu

fix-1157

bug-page-vie-privée

use-watername

fix-1176

carte-du-onde

fix-1166

page-news

tchou

add-icon-alias

vitesse

photon-opensearch

private-in-italic

bâtiments-3D

merge-categories

handle-private-highway

improve-findcategory

injection-pmtiles

fix-1128

check-icons

update-readme

category-rugby

bug-position-carte-vs-lieu

limites-vitesse

improve-find-category-from-tags-v2

integrer

document-osmfeature

translate-document

velo-fin-aout-25

velo-ete-2025

moins-de-maptiler

planet-pmtiles-tilemaker

fix-1048

panoramax-leger

plume

horaires-i18n

planet

maj-deps-juillet-2025

improve-find-category-from-tags

bug-encart

interactions-encart-mobile

clic-on-station

fix-1028

fix-1014

edit

add-big-rivers

add-shortbread-tiles

handle-deleted-osm-elements-v2

codeberg

improve-useEffect

immoloc

update-dashboard

analyze

terradraw

sombre2

calendar

poi-sprites-using-svg

style-sombre

sprites

annuaire

motis-gtfs-attributes

next-styled-components

dev

cool2

cool

panoramax-angle

env-static-urls

borders-only

panoramax2

intelligence

debug-iles

nouveau-serveur

france-protomaps

turbo

fouss/cartes-#338

blog2

legislatives

presentation

pas-de-geoloc-si-hash-different

protomaps

vaul

carte-trains

citiz

maj-paquets

icones

ui2

bblog

voyagevoyage

test-image-origine

bugimagefin

searchParams

voyage

avion

perf

studio

piscine

journée

année

journée-transition

score-dès-acceuil

fix-space-problem

rebase-5août

correspondance

règles-séparées

indeps-acre

demo

indeps

indeps-cachés

indication-temps-plein

indeps-refactor-simulations

indeps-vue-comparaison

indeps-inversion

barème-continu

indeps-démo

textes-branches-sécu

engine-enum

2019

transport

publicodes

infrance-nouvelle-landing

better-translation

co2

simplifATMP

controle-SMIC

webi18n

bug-avantages-en-nature

revert-352-infrance-navigation-v2

laem-patch-1

robots.txt

improve-perf

iframe-bugfix

quickfixes

lib

irsimple

hiringinfrance.com

dégradé

refstate

johan

zrr

logo-interactif

revert-216-improve-tracking

redesign2

zrr2

fixtrad

correction-salaire-brut

petit-state

alien-consolidate

translate-ui

npm

alien-brut-net

translatable

différent

ir18

fix-typo

add-tracking

fix-collect-missing

défauts-partout

fix-double-unfold-three

fix-double-unfold

versement-transport

blocker-questions

Labels

Clear labels

🎨 Design

À propos de l'interface et de l'UX

💡 Idée

Nouvelle feature à intégrer

❓ question

Further information is requested

🌍 OSM

À propos de l'affichage et l'édition des éléments OSM

🌱 Amélioration

Amélioration/fix de l'existant

🐞 Bug

Something isn't working

📊 SEO

📋 Catégories

À propos des catégories de recherche et de l'annuaire

📖 Wiki

📝Avis

A propos de la fonction avis et de bluesky

📢 Oyez

📷 Photos

A propos des photos (sources, affichage)

🔍Recherche

A propos de la barre de recherche et de photon

Styles de la carte

Ticket à traiter rapidement

Cette branche doit être déployée comme une démo

dependencies

Pull requests that update a dependency file

javascript

Pull requests that update javascript code

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

cartes/web#947

No description provided.

Delete branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?